Put your prompts under CI

Prompt changes ship on vibes more often than anyone admits. Someone tweaks a system prompt, it looks better on two examples, and it goes out. Two weeks later a metric quietly drops and nobody connects the dots.

Make quality a build check

Nebula runs your eval suite on every pull request, stores a baseline, and fails the build when a case regresses beyond tolerance. You get a per-case delta in the PR, so the reviewer sees exactly what got better and what got worse.

It is the same discipline unit tests brought to code, applied to the fuzzy part of your system.